Optimization with FBPIC

Description

This examples shows how to perform a Bayesian optimization of a laser-plasma accelerator (LPA) using FBPIC simulations.

The LPA to be optimized is based on the LUX design [1] using ionization injection.

The objective function to optimize (maximize) is defined as

\[f = \frac{\sqrt{Q} E_{MED}}{100 E_{MAD}}\]

where \(Q\) is the beam charge, \(E_{MED}\) is the median energy, and \(E_{MAD}\) is the median absolute deviation energy spread. This objective is optimized by tuning 4 parameters:

• 'laser_scale': parameter in the range \([0.7, 1.05]\) that scales the energy of the laser, which for laser_scale=1 is \(2.56 \, \mathrm{J}\).

• 'z_foc': the focal position of the laser in millimetres, with range \([3, 7.5]\).

• 'mult': parameter in the range \([0.1, 1.5]\) that scales the concentration of nitrogen in the injection region.

• 'plasma_scale': parameter in the range \([0.6, 0.8]\) that scales the plasma density of all species.

The optimization is carried out using an AxSingleFidelityGenerator and a TemplateEvaluator. In this case, the function analyze_simulation that analyzes the output of each simulation is defined in a separate file analysis_script.py and imported into the main optimas script.

The example is set up to make use of a system of 4 GPUs, where each FBPIC simulation uses a single GPU and 4 simulations are carried out in parallel.

Scripts

The files needed to run the optimization should be located in a folder (named e.g., optimization) with the following structure:

optimization
├── run_example.py
├── template_simulation_script.py
└── analysis_script.py

The optimization is started by executing:

python run_example.py

The scripts needed to run this example can be seen below.

run_example.py (download)

"""Example Bayesian optimization of an LPA with FBPIC.

This example optimizes an LPA based on ionization injection using FBPIC
simulations.

The FBPIC simulations are performed using the template defined in the
`template_simulation_script.py` file.

In addition to the objective `f`, three additional parameters
are analyzed for each simulation and including in the optimization
history. The calculation of `f` and the additional parameters is performed
in the `analyze_simulation` function, which for convenience is here defined in
the `analysis_script.py` file.
"""

from optimas.generators import AxSingleFidelityGenerator
from optimas.evaluators import TemplateEvaluator
from optimas.explorations import Exploration
from gest_api.vocs import VOCS

from analysis_script import analyze_simulation


# Create VOCS object.
vocs = VOCS(
    variables={
        "laser_scale": [0.7, 1.05],
        "z_foc": [3.0, 7.5],
        "mult": [0.1, 1.5],
        "plasma_scale": [0.6, 0.8],
    },
    objectives={"f": "MAXIMIZE"},
    observables=["energy_med", "energy_mad", "charge"],
)


# Create generator.
gen = AxSingleFidelityGenerator(
    vocs=vocs,
    n_init=4,
)


# Create evaluator.
ev = TemplateEvaluator(
    sim_template="template_simulation_script.py",
    analysis_func=analyze_simulation,
    n_gpus=1,  # Use 1 GPU per simulation.
)


# Create exploration.
exp = Exploration(
    generator=gen, evaluator=ev, max_evals=100, sim_workers=4, run_async=True
)


# To safely perform exploration, run it in the block below (this is needed
# for some flavours of multiprocessing, namely spawn and forkserver)
if __name__ == "__main__":
    exp.run()

template_simulation_script.py (download)

"""Script for simulating an LPA with ionization injection with FBPIC."""

import numpy as np
from scipy.constants import c, e, m_e, m_p

from fbpic.main import Simulation
from fbpic.lpa_utils.laser import add_laser_pulse
from fbpic.lpa_utils.laser.laser_profiles import GaussianLaser
from fbpic.openpmd_diag import BoostedParticleDiagnostic
from fbpic.lpa_utils.boosted_frame import BoostConverter


def LUXlaser(
    energy_measured_joule,
    FWHM_x_um,
    FWHM_t_fs,
    lambda_laser=0.8,
    T_beamline=1,
    focus_factor=1,
    temporal_factor=1,
):
    """Get the laser peak intensity, a0, spot size and length."""
    energy_gauss_joule = (
        energy_measured_joule * T_beamline * focus_factor * temporal_factor
    )
    FWHMtoSigma = 2 * np.sqrt(2 * np.log(2))
    I_0 = energy_gauss_joule / (
        (np.sqrt(2 * np.pi)) ** 3
        * (FWHM_x_um * 1e-4 / FWHMtoSigma) ** 2
        * (FWHM_t_fs * 1e-15 / FWHMtoSigma)
    )
    a_0 = 8.5492970742069339e-10 * lambda_laser * np.sqrt(I_0)
    w_0 = FWHM_x_um * 2 / FWHMtoSigma
    c_tau = 2.0 * (FWHM_t_fs * 1e-15) / FWHMtoSigma * 299792458.0 * 1e6
    return I_0, a_0, w_0, c_tau


# Optimization parameters.
laser_scale = {{laser_scale}}
z_foc = {{z_foc}} * 1.0e-3
mult = {{mult}}
plasma_scale = {{plasma_scale}}
spot_scale = 1.0
t_scale = 1.0


# Plasma density profile parameters.
dens_z = np.linspace(0, 8e-3, 1000)
dens_h2 = plasma_scale * (
    5e23 * np.exp(-(((dens_z - 2.9e-3) / 1.0e-3) ** 2))
    + 6e23 * np.exp(-(((dens_z - 5.3e-3) / 1.7e-3) ** 4))
)
dens_n2 = plasma_scale * (0.5e23 * np.exp(-(((dens_z - 2.5e-3) / 0.5e-3) ** 2)))

dens_h2 = dens_h2 - (mult - 1) * dens_n2
dens_n2 = dens_n2 * mult


# The simulation box
Nz = 3200  # Number of gridpoints along z
zmax = 0.0e-6  # Right end of the simulation box (meters)
zmin = -80.0e-6  # Left end of the simulation box (meters)
Nr = 270  # Number of gridpoints along r
rmax = 135.0e-6  # Length of the box along r (meters)
Nm = 2  # Number of modes used


# Boost factor and converter.
gamma_boost = 5.0
boost = BoostConverter(gamma_boost)


# Maximum simulation length
Lmax = np.amax(dens_z + zmax - zmin)


# The simulation timestep (seconds)
dt = min(rmax / (2 * gamma_boost * Nr), (zmax - zmin) / Nz / c)


# Order of the field solver.
n_order = 32


# Whether to use the GPU.
use_cuda = True


# Plasma particles.
p_zmin = 0.0e-6  # Position of the beginning of the plasma (meters)
p_rmax = 135.0e-6

# Particles per cell (Hydrogen)
p_nz = 1  # Number of particles per cell along z
p_nr = 2  # Number of particles per cell along r
p_nt = 4  # Number of particles per cell along theta

# Particles per cell (Nitrogen)
p_nz_N = 2
p_nr_N = 2
p_nt_N = 4


# Laser parameters and profile.
e_l = 2.56 * laser_scale  # Energy Joule
w_l = 25.0 * spot_scale  # Width (intensity) FWHM mu
w0_flat = w_l * 1.0e-6 / 1.609  # Flat-top w0 from FWHM for N=100
tau_l = 34.0 * t_scale  # Duration (intensity) FWHM fs
I0, a0, w0, ctau = LUXlaser(e_l, w_l, tau_l)
w0 *= 1.0e-6
ctau *= 1.0e-6
z0 = -3 * ctau  # Laser centroid
laser_profile = GaussianLaser(a0, w0, ctau / c, z0, zf=z_foc)


# Plasma density functions.
def dens_func_H(z, r):
    """Hydrogen density function."""
    z_lab = z * gamma_boost
    return 2 * np.interp(z_lab, dens_z, dens_h2)


def dens_func_N(z, r):
    """Nitrogen density function."""
    z_lab = z * gamma_boost
    return 2 * np.interp(z_lab, dens_z, dens_n2)


def dens_func_e(z, r):
    """Electron density function."""
    return dens_func_H(z, r) + 5 * dens_func_N(z, r)


# The moving window
v_window = c


# Velocity of the Galilean frame (for suppression of the NCI)
v_comoving = -np.sqrt(gamma_boost**2 - 1.0) / gamma_boost * c


# The diagnostics
diag_period = 100  # Period of the diagnostics in number of timesteps
# Whether to write the fields in the lab frame
Ntot_snapshot_lab = 2
dt_snapshot_lab = (Lmax + (zmax - zmin)) / v_window / (Ntot_snapshot_lab - 1)
track_bunch = False  # Whether to tag and track the particles of the bunch


# The interaction length (meters) and time (seconds) of the simulation.
# (i.e. the time it takes for the moving window to slide across the plasma)
L_interact = Lmax  # the plasma length
T_interact = boost.interaction_time(L_interact, (zmax - zmin), v_window)


# Carrying out the simulation
if __name__ == "__main__":
    # Initialize the simulation object
    sim = Simulation(
        Nz,
        zmax,
        Nr,
        rmax,
        Nm,
        dt,
        zmin=zmin,
        boundaries={"z": "open", "r": "open"},
        initialize_ions=False,
        n_order=n_order,
        use_cuda=use_cuda,
        v_comoving=v_comoving,
        gamma_boost=gamma_boost,
        verbose_level=2,
        particle_shape="cubic",
        use_galilean=True,
    )

    # By default the simulation initializes an electron species (sim.ptcl[0])
    # Because we did not pass the arguments `n`, `p_nz`, `p_nr`, `p_nz`,
    # this electron species does not contain any macroparticles.
    # It is okay to just remove it from the list of species.
    sim.ptcl = []

    # Add the Helium ions (pre-ionized up to level 1),
    # the Nitrogen ions (pre-ionized up to level 5)
    # and the associated electrons (from the pre-ionized levels)
    atoms_N = sim.add_new_species(
        q=5 * e,
        m=14.0 * m_p,
        n=1,
        dens_func=dens_func_N,
        p_nz=p_nz_N,
        p_nr=p_nr_N,
        p_nt=p_nt_N,
        p_zmin=p_zmin,
        p_rmax=p_rmax,
    )
    atoms_H = sim.add_new_species(
        q=e,
        m=1 * m_p,
        n=1,
        dens_func=dens_func_H,
        p_nz=p_nz,
        p_nr=p_nr,
        p_nt=p_nt,
        p_zmin=p_zmin,
        p_rmax=p_rmax,
    )
    elec = sim.add_new_species(
        q=-e,
        m=m_e,
        n=1,
        dens_func=dens_func_e,
        p_nz=p_nz,
        p_nr=p_nr,
        p_nt=p_nt,
        p_zmin=p_zmin,
        p_rmax=p_rmax,
    )

    # Activate ionization of N ions (for levels above 5).
    # Store the created electrons in a new dedicated electron species that
    # does not contain any macroparticles initially
    elec_from_N = sim.add_new_species(q=-e, m=m_e)
    atoms_N.make_ionizable("N", target_species=elec_from_N, level_start=5)

    # Add a laser to the fields of the simulation
    add_laser_pulse(
        sim,
        laser_profile,
        gamma_boost=gamma_boost,
        method="antenna",
        z0_antenna=0,
    )

    # Convert parameter to boosted frame
    (v_window,) = boost.velocity([v_window])

    # Configure the moving window
    sim.set_moving_window(v=v_window)

    # Add a diagnostics
    write_dir = "diags"
    sim.diags = [
        BoostedParticleDiagnostic(
            zmin,
            zmax,
            c,
            dt_snapshot_lab,
            Ntot_snapshot_lab,
            gamma_boost,
            diag_period,
            sim.fld,
            species={"electrons from N": elec_from_N},
            comm=sim.comm,
            write_dir=write_dir,
        )
    ]

    # Remove step 0 outputs
    sim.diags[0].snapshots.pop(0)

    # Calculate number of simulation steps to perform.
    N_step = int(T_interact / sim.dt)

    # Run the simulation
    sim.step(N_step)

analysis_script.py (download)

"""Defines the analysis function that runs after the simulation."""

import os
from openpmd_viewer.addons import LpaDiagnostics
import numpy as np
from scipy.constants import e


def analyze_simulation(simulation_directory, output_params):
    """Analyze the simulation output.

    This method analyzes the output generated by the simulation to
    obtain the value of the optimization objective and other analyzed
    parameters, if specified. The value of these parameters has to be
    given to the `output_params` dictionary.

    Parameters
    ----------
    simulation_directory : str
        Path to the simulation folder where the output was generated.
    output_params : dict
        Dictionary where the value of the objectives and analyzed parameters
        will be stored. There is one entry per parameter, where the key
        is the name of the parameter given by the user.

    Returns
    -------
    dict
        The `output_params` dictionary with the results from the analysis.
    """
    # Open simulation diagnostics.
    d = LpaDiagnostics(os.path.join(simulation_directory, "diags/hdf5"))

    # Get beam particles with `u_z >= 10` and transverse offset no larger than
    # 15 µm in `x` and `y`.
    uz, w = d.get_particle(
        ["uz", "w"],
        iteration=1,
        select={"uz": [10, None], "x": [-15e-6, 15e-6], "y": [-15e-6, 15e-6]},
    )

    # Convert charge to pC.
    q = w.sum() * e * 1e12

    # Analyze distribution and fill in the output data.
    if len(w) < 2:  # Need at least 2 particles to calculate energy spread
        output_params["f"] = 0
    else:
        med, mad = weighted_mad(uz / 2, w)
        output_params["f"] = np.sqrt(q) * med / mad / 100
        output_params["charge"] = q
        output_params["energy_med"] = med
        output_params["energy_mad"] = mad

    return output_params


def weighted_median(data, weights):
    """Compute the weighted quantile of a 1D numpy array.

    Parameters
    ----------
    data : ndarray
        Input array (one dimension).
    weights : ndarray
        Array with the weights of the same size of `data`.
    quantile : float
        Quantile to compute. It must have a value between 0 and 1.

    Returns
    -------
    quantile_1D : float
        The output value.

    """
    quantile = 0.5
    # Check the data
    if not isinstance(data, np.matrix):
        data = np.asarray(data)
    if not isinstance(weights, np.matrix):
        weights = np.asarray(weights)
    nd = data.ndim
    if nd != 1:
        raise TypeError("data must be a one dimensional array")
    ndw = weights.ndim
    if ndw != 1:
        raise TypeError("weights must be a one dimensional array")
    if data.shape != weights.shape:
        raise TypeError("the length of data and weights must be the same")
    if (quantile > 1.0) or (quantile < 0.0):
        raise ValueError("quantile must have a value between 0. and 1.")
    # Sort the data
    ind_sorted = np.argsort(data)
    sorted_data = data[ind_sorted]
    sorted_weights = weights[ind_sorted]
    # Compute the auxiliary arrays
    Sn = np.cumsum(sorted_weights)
    # TODO: Check that the weights do not sum zero
    # assert Sn != 0, "The sum of the weights must not be zero"
    Pn = (Sn - 0.5 * sorted_weights) / Sn[-1]
    # Get the value of the weighted median
    return np.interp(quantile, Pn, sorted_data)


def weighted_mad(x, w):
    """Calculate weighted median absolute deviation."""
    med = weighted_median(x, w)
    mad = weighted_median(np.abs(x - med), w)
    return med, mad

References

[1]

Sören Jalas, Manuel Kirchen, Philipp Messner, Paul Winkler, Lars Hübner, Julian Dirkwinkel, Matthias Schnepp, Remi Lehe, and Andreas R. Maier Phys. Rev. Lett. 126, 104801 (2021)